Improving Accuracy of IC Surface Defects Detection via Enhanced-CycleGAN Data Augmentation

An important step in integrated circuit (IC) manufacturing is inspection of the chip surface for defects. In practice, there exists a data imbalance problem associated with IC surface defect images which affects the detection performance of a deep learning-based detection model. In this paper, this data imbalance problem is addressed via generating synthetic IC surface defect images by the generative network of Enhanced-CycleGAN. The main contribution of this work involves exhibiting that such a data augmentation enables defective IC surfaces to be detected at higher detection accuracies compared with the situation when no data augmentation is used. First, the generated synthetic images by the Enhanced-CycleGAN generative network are compared with the synthetic generative networks of GAN, VAE, Diffusion, and CycleGAN as well as with the traditional image processing data augmentation. The closeness of synthetic images to real images is assessed based on five similarity metrics of Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), Visual Information Fidelity (VIF), Fréchet Inception Distance (FID), and Kernel Inception Distance (KID). The assessment conducted based on these metrics indicates that the synthetic IC surface defect images generated by Enhanced-CycleGAN resemble closer to the real IC surface defect images compared to the other generative networks. Second, the impact of the data augmentation by the Enhanced-CycleGAN generative network on the detection of IC surface defects is examined by considering three typical and commonly used detection models of VGG16, ResNet50, and YOLO11. It is shown that our Enhanced-CycleGAN data augmentation improves the detection performance across all the three detection models.